PhytoBench-Knowledge

Evaluation benchmark for the Knowledge Agent of the Phytomni multi-agent system. The benchmark covers 65 plant species and is split into two complementary sub-benchmarks distinguished by the scene field:

• ID (700 pairs) — gene name identification from functional descriptions.
• Trace (300 pairs) — detailed knowledge tracing in the scientific literature.

Construction

ID sub-benchmark. 700 Q&A pairs targeting gene name identification. Gene IDs and symbols were drawn from leading species-specific resources, including TAIR, funRiceGenes, WheatOmics, MaizeGDB, SoyMD, and funPlantGenes. Pairs that overlapped with the Phytomni-SFT corpus were excluded; the remaining pool was manually curated to ensure correctness.

Trace sub-benchmark. 300 expert-curated Q&A pairs assessing detailed knowledge tracing. Ten domain specialists — covering Omics Data Integration, Abiotic Stress Adaptation, Gene Editing, Synthetic Biology, Trait Mechanisms, Epigenetics, Developmental Biology, Plant–Microbe Interactions, Biotic Stress Resistance, and Smart Design — each selected 3–5 non-Open-Access papers from the past decade and authored questions meeting three criteria: (1) information absent from the abstract, (2) does not require figure/table interpretation, and (3) represents a unique finding or methodology. SFT-overlap pairs were excluded; the remaining items underwent two-stage manual and expert review.

Schema

Field	Type	Description
id	string	Stable identifier for the Q&A pair.
species	string	Source species (one of 65 plant species).
query	string	The question presented to the agent.
answer	string	Reference answer. For ID, a gene ID or symbol (or pipe-separated synonyms). For Trace, a free-text expert answer.
scene	string	One of ID or Trace. Selects the sub-benchmark and its evaluation protocol.
reference	string	Citation or source pointer for the supporting evidence; populated for the Trace subset.

How to use

from datasets import load_dataset

ds = load_dataset("Phytomni/PhytoBench-Knowledge", split="test")

id_set    = ds.filter(lambda r: r["scene"] == "ID")     # 700 rows
trace_set = ds.filter(lambda r: r["scene"] == "Trace")  # 300 rows

Evaluation protocol. ID is scored by exact match via regular expressions: a generated gene ID or symbol is correct if it matches one of the pipe-separated synonyms in answer. Trace is scored by BLEU-4 between the generated text and the expert reference in answer.

Citation

@article{phytomni2026,
  title  = {Phytomni: An agentic AI accelerating plant research from discovery to design},
  author = {Phytomni Team},
  year   = {2026},
  note   = {Manuscript in preparation; citation TBD until publication.}
}

Links

• Collection: https://huggingface.co/collections/Phytomni/phytobench
• Project: https://github.com/Phytomni/Phytomni
• Agent code: https://github.com/Phytomni/Phytomni-Bot

License: GPL-3.0

Version

v0.1.0